Design and Fabrication of the Sensor

5.2.1 Design and Sensing Principle

A quad-unit force sensing cell consisting of 4 internal individual force units (𝑆₁ , 𝑆₂, 𝑆₃, 𝑆₄), that are axisymmetrically distributed is utilized as shown in Figure 5-2. Each force unit has two electrodes. The electrodes are patterned in such a way that the four units lie symmetrically on the axes of a Cartesian coordinate system with the center of the eight electrodes sitting at the origin. The electrolyte is cured on the surface of a soft semi- ellipsoid. This soft semi-ellipsoid is assembled over the electrodes. If only normal force is applied, as shown in Figure 5-2 (a) and (b), the contact area between the electrodes and electrolyte increases, resulting in an increase in capacitance of all four sensing units. If both normal and shear forces are applied (Figure 5-2(c)), the center of the contact area shifts accordingly depending on the direction of the shear force, with the readings from the four force units being different.

Figure 5-2 Working mechanism of the supercapacitve normal and shear sensor (a) no force is applied, (b) normal force is applied, (c) both normal and shear forces are applied.

The physical structure of the sensor is shown in Figure 5-3. The 4 pairs of electrodes are patterned in a circle on a soft substrate. The top chamber is made of both soft and hard material. It consists of a hard cap, a hard mesh, a soft semi-ellipsoid substrate to host the solid electrolyte and a soft side wall. The hard cap (Figure 5-3(a)) is designed for better receiving the force applied from the top. A hard mesh (or skeleton) is printed and embedded inside the soft side wall to increase the elasticity of the sensor (Figure 5-3(c) and (d)). The solid electrolyte is cured on the surface of the soft semi-ellipsoid, which can deform under the applied force. The 4 capacitance readings (𝐶1, 𝐶2, 𝐶3, 𝐶4) of the 4 sensing units are used

to estimate the normal force and shear force applied on the sensing cell.

Figure 5-3 Schematic of the normal and shear force sensor (a) top chamber with the solid electrolyte, (b) bottom electrodes patterned on a soft substrate, (c) three quarter section view of the

5.2.2 Fabrication of the Sensor

The bottom electrodes are fabricated on a flexible substrate, namely a polyimide film (25 µm, Pyralux Copper Kapton Laminate, Dupont), using MEMS fabrication technologies, as shown in Figure 5-4(a). First, a thin layer of titanium (20nm) is sputtered on the polyimide film (PI) for enhancing adhesion, on top of which a gold layer (200nm) is sputtered. Then, the 8 electrodes are patterned in a circle using photolithography and wet etching. The as-fabricated electrode is shown in Figure 5-4(b).

Figure 5-4 (a) Fabrication process of the bottom electrode. (b) The electrodes patterned on PI substrate

The paper-based solid electrolyte is cured on top of an arch structure inside a 3D printed chamber as shown in Figure 5-5. The fabrication process starts from cutting the filter paper into a circular shape followed by pre-shaping it into a dome shape. The arch was then transferred to the top of the soft arch of the 3D printed chamber. An ionic gel was brushed onto the paper, that functionalizes the entire thickness of the paper and results in a solid- state electrolyte. After that, the combination was put under UV light in a glovebox for 1 min. A clear film (150um) sticking to the 3D printed part is obtained. The 3D printed chamber with electrolyte is shown in Figure 5-5(c). The part is made using multi-material 3D printing technology, combining materials of different properties in a single part. The structure contains a soft side wall (Stratasys, Agilusclear), a hard cap, a soft ellipsoid and a hard mesh (Stratasys, Verowhite). The cap is printed for the purpose of better receiving forces applied on top, while a hard mesh (or skeleton) is embedded inside the soft body and designed to improve the elasticity of the structure.

Figure 5-5 (a) Fabrication of the electrolyte. (b) 3D printed chamber and the pre-shaped filter paper. (c) Solid electrolyte cured inside the soft chamber

The soft chamber with the cured electrolytes was then assembled over the electrodes using sealing glue (Loctite, Waterproof Sealant). Connecting wires (40 AWG, enamel wires) are soldered onto the electrical contact pads on the PI substrate. The other ends of the wires are connected to 4 jumper wires for easy connection to the measurement equipment. The assembled sensor with 4 supercapacitive sensing units is shown in Figure 5-6.

Figure 5-6 Photograph of the assembled normal and shear force sensor.

5.2.3 Learning of the Sensor Response

In order to provide accurate estimates of the shear and normal forces acting on the top of the sensor, an accurate sensor model must be used to fully map from individual force unit readings to the applied normal and shear forces. For a linear sensor with axisymmetric stiffness, the normal force can be obtained from the average of the four force units as

𝐹_𝑧 =1

4(𝐾𝑛1𝐶1 + 𝐾𝑛2𝐶2+ 𝐾𝑛3𝐶3+ 𝐾𝑛4𝐶4) (5-1) where 𝐾𝑛1, 𝐾𝑛2, 𝐾𝑛3 and 𝐾𝑛4 are the calibration coefficients between normal force and

capacitance of each sensor.

The shear force along the x axis can be obtained from:

𝐹𝑥 = |(𝐾𝑠3𝐶3 + 𝐾𝑠4𝐶4) − (𝐾𝑠1𝐶1+ 𝐾𝑠2𝐶2)| (5-2)

While the shear force along the y axis can be obtained from:

𝐹_𝑦 = |(𝐾_𝑠1𝐶₁+ 𝐾_𝑠4𝐶₄) − (𝐾_𝑠3𝐶₃+ 𝐾_𝑠2𝐶₂)| (5-3) where 𝐾𝑠1, 𝐾𝑠2, 𝐾𝑠3 and 𝐾𝑠4 are the calibration coefficients between shear force and

capacitance of each sensor.

However, the fabricated shear and normal force sensor response is highly non-linear and complex, with forces being exerted from 3 degrees of freedom and the sensors not being axis-symmetric due to significant manufacturing imperfections. That can make accurately building an analytical sensor model very difficult in practice. Therefore, a generalized regression neural network [126]–[128] was used in order to obtain a model matching the experimental data as well as possible. A two-layer feed-forward network with sigmoid hidden neurons and linear output neurons, is designed to fit this multi-dimensional mapping problem as shown in Figure 5-7. The capacitance readings of the four sensing units are fed into the neural network as the inputs, while the 3D forces (measured by a load cell during training) are used as the outputs of the neural network. The network is trained with Levenberg-Marquardt backpropagation algorithm [129].

Figure 5-7 Schematic of the neural network